Younes and I were chatting yesterday and since he is now collecting real participant data, we recognized a missing component to AutoRA - i.e., processing of data. For example, outlier analysis. With Younes' current task, he has code that removes any participant under a certain accuracy. Right now, AutoRA has no capability to process data in this way, and so I think we could add a new component that does this.
Essentially, this component operates on the experimental data to either remove data (e.g., in terms of outlier analysis), or transform data (e.g., filtering data, normalize data, or whatever might be applicable to the task), or potentially something else (I am sure this can take many forms). This would at least need a new field in the state - something like rejected data - where it can keep track on how it manipulated things. For example, a rejected outlier based on SD would be moved from experimental data to rejected data because we would like to keep track of what's rejected but also it may not be rejected in the next cycle if we accumulated more data and suddenly it's not as far from the mean. For transformations, it might be necessary to keep a new field - unprocessed data - that contains all raw data unprocessed.
These are just first thoughts, so it can take many forms but I wanted to share.
Younes and I were chatting yesterday and since he is now collecting real participant data, we recognized a missing component to AutoRA - i.e., processing of data. For example, outlier analysis. With Younes' current task, he has code that removes any participant under a certain accuracy. Right now, AutoRA has no capability to process data in this way, and so I think we could add a new component that does this.
Essentially, this component operates on the experimental data to either remove data (e.g., in terms of outlier analysis), or transform data (e.g., filtering data, normalize data, or whatever might be applicable to the task), or potentially something else (I am sure this can take many forms). This would at least need a new field in the state - something like rejected data - where it can keep track on how it manipulated things. For example, a rejected outlier based on SD would be moved from experimental data to rejected data because we would like to keep track of what's rejected but also it may not be rejected in the next cycle if we accumulated more data and suddenly it's not as far from the mean. For transformations, it might be necessary to keep a new field - unprocessed data - that contains all raw data unprocessed.
These are just first thoughts, so it can take many forms but I wanted to share.